Question: Does earthquake shaking change the discharge of United States streams in rain-free periods, and does the change have a threshold or a dose response? Approach: Earthquake time is independent of weather. We compared each gauge after an earthquake with the same gauge in the same season of other years. We used 94 earthquakes (M >= 5.5, 1980 to 1990), USGS daily discharge, GHCN-Daily rain data, and seismic energy density e as the exposure. A plan and two amendments were written before outcomes were computed. Filters kept 1,276 gauge-event pairs from 22 earthquakes and 711 gauges. Findings: At e >= 0.1 J/m3, the literature threshold, the median discharge change on days +1 to +5 was +2.2 percent (124 pairs, 13 events; placebo p = 0.23). The slope of the median effect against log10 e was -0.009 per decade (95 percent interval -0.113 to +0.054; placebo p = 0.77). For e of 0.01 to 0.1 J/m3 the 95 percent interval was -0.9 to +2.1 percent. For e of 0.1 to 1 J/m3 it was 0.0 to +7.6 percent. A test for a larger response above 1 J/m3 was inconclusive (19 pairs, 7 events). No retained pair reached 10 cm/s peak ground velocity, so we could not test the binary response reported by Mohr et al. (2018). Limitations: Downloads stopped at a time limit, so the study covers 1980 to 1990 and only 20 ShakeMap grids. Rain and interference filters removed most pairs near strong earthquakes. A median test has low power when only a minority of gauges respond. Exploratory results suggest a response at Loma Prieta 1989 (17 pairs, median +11.5 percent), but this is a post-hoc look at one event.
Filters: key FXiq8nj2Ene7ql7CaSCnG0UTbNdSfzsMU678zYa0fmY ×
Question: Does earthquake shaking raise stream discharge more often than chance, and at which exposure does the excess start? Approach: We tested all available United States gauges near 385 mainshocks (M 5 or more, 1950 to 2024), not only streams with reported changes. We used USGS daily discharge, ComCat earthquakes, seismic energy density e, and ShakeMap peak ground velocity. A pair was eligible if the gauge was in a dry state. There were 25,004 eligible pairs. Each pair was compared with the same gauge in the same season of 30 other years. We wrote the plan before we computed outcomes. Findings: At e of 1 to 10 J/m3, 44.8 percent of eligible gauges rose by 20 percent or more, against 21.0 percent in placebo windows (excess +23.8 points, 95 percent interval +15.1 to +31.9, Holm p = 0.005). At e of 10 J/m3 or more the excess was +27.3 points. At 0.1 to 1 J/m3 and 0.01 to 0.1 J/m3 the excess was +6.9 and +5.1 points (Holm p = 0.087). These lower bins are not established. With peak ground velocity the excess was +8.7 points at 3 to 10 cm/s and +31.9 points at 10 to 30 cm/s. A smooth logistic curve predicted held-out events better than a step, so we do not claim a sharp threshold. Decreases did not increase. At e of 1 J/m3 or more, the mean excess in ln discharge grew to +0.31 by day 20 and did not decay by day 30, so we could not estimate a recovery time. Post-hoc checks did not remove the excess: same-year pre-event windows, ring contrast within events (+25.7 points), and June to October events (+20.8 points). Limitations: We had no rain data. Five events supply 59 percent of responding pairs at e of 1 J/m3 or more, and without the top ten events the excess is +8.6 points. The result applies only to dry pre-event states. Epicentral energy density ignores rupture geometry. The planned positive control failed the planned eligibility rule, because NWIS rounds low flows; a post-hoc rule restored it.
We tested if monthly changes in water mass on land change the rate of shallow earthquakes. We used GRACE and GRACE-FO mascons and the USGS ComCat catalog for 2002 to 2026. The primary sample has 1,942 declustered events of M 5.0 or more at depth 40 km or less, in 623 land cells of 2 degrees. A Poisson model with cell and year fixed effects separated seasonal and non-seasonal load, and load rate. The analysis plan was written before the first fit. No primary term passed the planned test (smallest Holm p = 0.31). Seasonal load gave -6.8% per kPa (95% bootstrap interval -13.5 to -1.5, null p = 0.078). The bootstrap interval excludes zero but the null test does not reject zero, so we do not report a detection. Non-seasonal load gave +2.6% per kPa (-5.1 to +11.4). The seasonal estimate changed with the magnitude floor (-2.0 at M 4.5, +0.9 at M 5.5). The design can detect approximately 10 to 12% per kPa, so smaller effects are not excluded. Limits: the test covers land cells only, GRACE resolution is approximately 300 km, regional opposite signs can cancel, and lags were not tested. An injection test showed possible attenuation of the seasonal rate term, and a longer repeat stopped at a time limit before it could resolve this.
Question: does the input anatomy of one excitatory cortical neuron predict how reliably it responds to a repeated movie and how strongly it follows the population? Does the relation found in one mouse hold in a second mouse? Approach: we used calcium imaging and electron microscopy of the same neurons in the MICrONS mouse (12,685 neurons) and the V1 Deep Dive mouse (179 neurons with complete input tables). For each neuron we counted soma, spine and shaft input synapses. We fit one model in each mouse with fixed effects for session, imaging plane and cell type, and controls for position, nucleus volume and imaging noise. We used a cluster bootstrap. We wrote and hashed the plan before we computed these relations. A relation counted as replicated only if it had Holm-adjusted significance with the same sign in both mice. Results: no relation met the replication rule. All six V1DD Holm p values were 0.86 or more, and V1DD intervals were wide. In MICrONS, spine input count predicted higher population coupling (partial slope 0.150, 95% interval 0.120 to 0.185, Holm p 0.006). The V1DD estimate was 0.116 (interval -0.045 to 0.279). In MICrONS, soma input count did not predict reliability (slope -0.007, interval -0.030 to 0.013) or coupling (-0.008, interval -0.031 to 0.014). Thus perisomatic input count had no clear relation to these two properties. Shaft input count predicted higher reliability in MICrONS (0.103, interval 0.074 to 0.132). A model fit in V1DD predicted held-out MICrONS coupling with r = 0.060 (interval 0.039 to 0.080, Holm p 0.002), which is about 0.3% of the residual variance. The reverse direction had r = 0.122 (interval -0.043 to 0.264, Holm p 0.33). Reliability did not transfer. The spine-coupling slope was negative in 488 MICrONS L5-ET cells. It was positive in the IT types. Limitations: two animals, a small V1DD sample after the removal of partial input tables, observational data, labels that are proxies for input type, input counts that correlate with one another, and different movies and traces between the mice. A plan amendment, made before any test, removed V1DD neurons with partial input tables. We supply code, tables, figures, and data hashes.